---

README / Overview of the Autonomous Deforestation & Wildfire Detection System

Project Summary

This project simulates an autonomous wildfire detection and containment system using drones, satellite imagery, and AI. It uses reinforcement learning (PPO) to coordinate a swarm of drones to prevent the spread of forest fires.

---

System Components

1️⃣ SatelliteIngestionAgent

• Purpose: Loads satellite imagery (Sentinel/MODIS) to detect vegetation health and thermal anomalies.

• Key Steps:

  • Loads .tif files for NDVI (Normalized Difference Vegetation Index) and thermal data.
  • If real data not available → simulates data randomly.
  • Preprocesses images for the model (resize + normalize).

Key terms:

• NDVI: Measures vegetation health. Range [-1, 1].
• Thermal anomalies: Areas that are unusually hot → could indicate fire.

---

2️⃣ ThreatDetectionAgent

• Uses a ResNet18 model for feature extraction from satellite images.
• Combines features with NDVI and thermal data.
• Uses Isolation Forest to detect anomalies (fire-prone areas).

Key terms:

• ResNet18: Neural network for feature extraction from images.
• Isolation Forest: ML model to detect anomalies (outliers).

---

3️⃣ DecisionAgent

• Scores risk based on threat detection.

• Computes confidence for each risk.

• Decides action:

  • deploy_drones → if high risk & high confidence
  • monitor → if low risk

Key terms:

• Confidence: Likelihood that the detected threat is real.
• Action escalation: Deciding whether drones should intervene.

---

4️⃣ DroneSwarmAgent

• Controls multiple drones.
• Each drone has a role: scout, containment, monitor.
• Drones move using a simple A pathfinding* logic to reach targets.

Key terms:

• Scout: Explore unexplored areas.
• Containment: Fight fire edges.
• Monitor: Stay put or observe.
• Battery: Energy consumed per move.

---

5️⃣ ContainmentAgent

• Plans firebreaks to prevent fire spread.
• Applies retardant to reduce fuel load and slow down fire.

Key terms:

• Firebreaks: Paths cleared to stop fire.
• Fuel load: Amount of combustible material.

---

6️⃣ ForestDigitalTwin

• Digital simulation of the forest as a grid graph.

• Nodes = forest patches, edges = adjacency.

• Attributes per node:

  • on_fire → is this patch burning
  • fuel → amount of flammable material
  • spread_prob → probability fire spreads to this patch
  • visited → if drone has visited

Key terms:

• Digital twin: Virtual model of a real-world system.
• Wind factor & terrain cost: Affect fire spread probability.

---

7️⃣ MultiDroneEnv (Gymnasium Environment)

• Custom RL environment for training drones.
• Action space: MultiDiscrete (up/down/left/right/stay) per drone.
• Observation space: NDVI + fire map (2 channels).
• Reward: Encourages putting out fire and minimizing energy.

Key terms:

• Reward: Positive points for extinguishing fire, negative for energy waste.
• Episode done: Drones depleted or fire fully contained.

---

Simulation Flow

1. System Initialization

   • All agents (Satellite, Threat, Decision, Drone, Containment) are initialized.
   • Digital twin created for forest.

2. Data Ingestion

   • Load satellite tiles and NDVI maps.
   • Simulate if data is missing.

3. Threat Detection

   • Detect potential fire spots via ML model.
   • Compute risk and confidence.
   • Decide actions (deploy drones or monitor).

4. Drone Movement

   • Drones move according to roles.
   • Fire updates in digital twin based on spread rules.

5. Reinforcement Learning

   • PPO trains drones for optimal containment policy.
   • DummyVecEnv wraps the environment for Stable-Baselines3.
   • Stop training when reward threshold reached (example: 50).

6. Containment & Metrics

   • Plan firebreaks & apply retardant.
   • Update fire spread post-drone intervention.
   • Compute metrics.

---

Metrics Explained

Metric	Meaning
Latency	Time system took to run the simulation. Low = faster response.
Area Saved	How many initial fire patches were successfully contained.
Time to Containment	Time elapsed until fire is mostly contained.
Energy Consumed	Sum of drone battery usage during simulation.
CO2 Avoided	Environmental impact: more fire contained → more CO₂ saved.

Note: If Area Saved = 0 and CO2 Avoided = 0, it means drones didn’t successfully contain any fires in this run.

---

Visualization

• NDVI Map → background of forest health.
• Drones → red dots.
• Active fires → orange dots.
• Updates dynamically each step in simulation.

---

How to Use

1. Install dependencies: pip install torch torchvision rasterio stable-baselines3 gymnasium matplotlib networkx streamlit
2. Run Streamlit app:

streamlit run wildfire_sim.py

3. Press "Start Simulation" → Watch drones react, fires spread, and metrics update.
4. Metrics logged in metrics.csv for analysis.

---

Key Features

• Integrates satellite imagery, ML anomaly detection, digital twin modeling, and drone swarm control.
• Real-time metric reporting: area saved, CO₂ avoided, energy used.
• RL-based drone coordination using PPO.

---